A Recurrent Neural Network Approach to Audio Segmentation for Broadcast Domain Data

This paper presents a new approach for automatic audio segmentation based on Recurrent Neural Networks. Our system takes advantage of the capability of Bidirectional Long Short Term Memory Networks (BLSTM) for modeling temporal dy-namics of the input signals. The DNN is complemented by a resegmentation module, gaining long-term stability by means of the tied-state concept in Hidden Markov Models. Further-more, feature exploration has been performed to best represent the information in the input data. The acoustic features that have been included are spectral log-filter-bank energies and musical features such as chroma. This new approach has been evaluated with the Albayz´ın 2010 audio segmentation evaluation dataset. The evaluation requires to differentiate five audio conditions: music, speech, speech with music, speech with noise and others. Competitive results were obtained, achieving a relative improvement of 15.75% compared to the best results found in the literature for this database.